Improving realism in patient-specific abdominal Ultrasound simulation using CycleGANs

S. VITALE; J. I. ORLANDO; E. IARUSSI; I. LARRABIDE

Congreso; Computer Assisted Radiology and Surgery Congress; 2019

Purpose: In this paper we propose to apply generative adversarial neu-ral networks trained with a cycle-consistency loss, or CycleGANs, to improve re-alism in ultrasound (US) simulation from Computed Tomography (CT) scans.Methods: A ray-casting US simulation approach is used to generate intermedi-ate synthetic images from abdominal CT scans. Then, an unpaired set of thesesynthetic and real US images is used to train CycleGANs with two alternativearchitectures for the generator, a U-Net and a ResNet. These networks are finallyused to translate ray-casting based simulations into more realistic synthetic USimages.Results: Our approach was evaluated both qualitatively and quantitatively. A userstudy performed by 21 experts in US imaging shows that both networks signif-icantly improve realism with respect to the original ray-casting algorithm (p 0.0001), with the ResNet model performing better than the U-Net (p 0.0001).Conclusion: Applying CycleGANs allows to obtain better synthetic US images ofthe abdomen. These results can contribute to reduce the gap between artificiallygenerated and real US scans, which might positively impact in applications suchas semi-supervised training of machine learning algorithms and low-cost trainingof medical doctors and radiologists in US image interpretation.